{ "id": 255607, "title": "Apperently the author is more important then the code", "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/255607", "author_name": "", "post_date": "2021-07-28T10:49:45.302263700Z", "votes": 40, "comment_count": 17, "views": 0, "content": "

Hi,

\n

I am asking myself why this notebook:

\n

Brain Tumor

\n

got 26 upvote so far

\n

and this notebook:

\n

Brain Tumor very basic inference

\n

got 10 upvote so far,

\n

even thought the first one was copied from the second one

\n

and the author of the first one, that copied not even gave a upvote

", "messages": [ { "id": "1402600", "postDate": "07/28/2021 10:49:45", "content": "

Hi,

\n

I am asking myself why this notebook:

\n

Brain Tumor

\n

got 26 upvote so far

\n

and this notebook:

\n

Brain Tumor very basic inference

\n

got 10 upvote so far,

\n

even thought the first one was copied from the second one

\n

and the author of the first one, that copied not even gave a upvote

", "rawMarkdown": "Hi,\n\nI am asking myself why this notebook:\n\n[Brain Tumor](https://www.kaggle.com/blade001/brain-tumor-code)\n\ngot 26 upvote so far\n\nand this notebook:\n\n[Brain Tumor very basic inference](https://www.kaggle.com/lucamtb/brain-tumor-very-basice-inference)\n\ngot 10 upvote so far, \n\neven thought the first one was copied from the second one\n\nand the author of the first one, that copied not even gave a upvote", "votes": null }, { "id": "1402607", "postDate": "07/28/2021 10:55:47", "content": "

I think that posts were no respect for the original post. Thanks Luca !

", "rawMarkdown": "I think that posts were no respect for the original post. Thanks Luca !", "votes": null }, { "id": "1402636", "postDate": "07/28/2021 11:39:46", "content": "

Dear @lucamtb

\n

On kaggle there is certainly a Matthew effect where higher ranked kagglers will procure, for better or for worse, more attention. However, in this case I would suggest the principal motive is the slight improvement in the LB score, which attracts some competition participants like moths to a flame. Indeed many such participants in competitions who do not undertake their own work are often caught out towards the end of many competitions by blindly submitting high scoring blended ensembles that turn out to be horrendously overfitted.
\nIt is certainly poor form to not have cited your original notebook when forking, more so given that the only difference was the slight change of one parameter.

\n

I hope you get the upvotes you deserve.

\n

All the best,
\ncarl

", "rawMarkdown": "Dear @lucamtb \n\nOn kaggle there is certainly a [Matthew effect](https://en.wikipedia.org/wiki/Matthew_effect) where higher ranked kagglers will procure, for better or for worse, more attention. However, in this case I would suggest the principal motive is the slight improvement in the LB score, which attracts some competition participants like moths to a flame. Indeed many such participants in competitions who do not undertake their own work are often caught out towards the end of many competitions by blindly submitting high scoring blended ensembles that turn out to be horrendously overfitted.\nIt is certainly poor form to not have cited your original notebook when forking, more so given that the only difference was the slight change of one parameter. \n\nI hope you get the upvotes you deserve.\n\nAll the best,\ncarl", "votes": null }, { "id": "1403244", "postDate": "07/29/2021 00:34:46", "content": "

I have noticed that too.

", "rawMarkdown": "I have noticed that too.", "votes": null }, { "id": "1403467", "postDate": "07/29/2021 06:34:59", "content": "

It happens most of time. At least you need to reference if you took some codes or inspired from another notebook. I hope people will learn it sooner or later.

", "rawMarkdown": "It happens most of time. At least you need to reference if you took some codes or inspired from another notebook. I hope people will learn it sooner or later.", "votes": null }, { "id": "1404555", "postDate": "07/30/2021 03:19:48", "content": "

I understand that the users filter notebooks on the score from the Code section and might have upvoted @blade001 's notebook from @lucamtb 's original work. The community would highly appreciate if the attribution of the original code author is added in the copied versions for better credibility.

\n

On the case of votes being cast - I think this is a noticeable issue with Kaggle as of now. Votes of notebooks don't pass to the original authors of the code work but that instance of the notebook.

", "rawMarkdown": "I understand that the users filter notebooks on the score from the Code section and might have upvoted @blade001 's notebook from @lucamtb 's original work. The community would highly appreciate if the attribution of the original code author is added in the copied versions for better credibility. \n\nOn the case of votes being cast - I think this is a noticeable issue with Kaggle as of now. Votes of notebooks don't pass to the original authors of the code work but that instance of the notebook.", "votes": null }, { "id": "1404585", "postDate": "07/30/2021 04:33:57", "content": "

That's sad and disappointing, but certainly happens, you should contact the person who fork your notebook and ask for the citation.

", "rawMarkdown": "That's sad and disappointing, but certainly happens, you should contact the person who fork your notebook and ask for the citation.", "votes": null }, { "id": "1404647", "postDate": "07/30/2021 05:59:52", "content": "

I believe in fostering collaboration and creating a place where we (as fellow ML life-long learners) can be great support for one another. Sharing my own personal rule, in the hopes that it will inspire others to do so as well.

\n

a) if you find content that has captured your attention, or you've learned something new, or was worth your while in some form or another, or you became captured reading it -- upvote it. because ultimately, that content contributed something to you even if in the smallest minutiae

\n

b) if a notebook is useful in some manner or reason, not even to mention if it is one you're going to \"fork\", then definitely upvote it (not just take it for granted that it says \"copied from…\" but to actually give that person credit where credit is due; afterall you came across their notebook and used it).

\n

Perhaps I am verbose about this, more so at the moment, because I happen to have stumbled across many threads that seem to present information that could truly jeopardize what this platform's reputation is known to stand for (learning, collaboration, knowledge sharing), and the reason why I have joined.

\n

Look forward to learning from all of you!

", "rawMarkdown": "I believe in fostering collaboration and creating a place where we (as fellow ML life-long learners) can be great support for one another. Sharing my own personal rule, in the hopes that it will inspire others to do so as well.\n\na) if you find content that has captured your attention, or you've learned something new, or was worth your while in some form or another, or you became captured reading it -- upvote it. because ultimately, that content contributed something to you even if in the smallest minutiae\n\nb) if a notebook is useful in some manner or reason, not even to mention if it is one you're going to \"fork\", then definitely upvote it (not just take it for granted that it says \"copied from...\" but to actually give that person credit where credit is due; afterall you came across their notebook and used it).\n\nPerhaps I am verbose about this, more so at the moment, because I happen to have stumbled across many threads that seem to present information that could truly jeopardize what this platform's reputation is known to stand for (learning, collaboration, knowledge sharing), and the reason why I have joined. \n\nLook forward to learning from all of you!", "votes": null }, { "id": "1405030", "postDate": "07/30/2021 12:36:19", "content": "

If the new post doesn't add any value and doesn't give attribution to the original post you should report it.

\n

Simple as that.

\n

ps: I gave you an upvote so maybe that'll help.

", "rawMarkdown": "If the new post doesn't add any value and doesn't give attribution to the original post you should report it. \n\nSimple as that.\n\nps: I gave you an upvote so maybe that'll help.", "votes": null }, { "id": "1406310", "postDate": "07/31/2021 16:35:04", "content": "

Well out of his three notebooks he copied two and got gold medals for that…

", "rawMarkdown": "Well out of his three notebooks he copied two and got gold medals for that...", "votes": null }, { "id": "1408336", "postDate": "08/02/2021 12:23:31", "content": "

I saw him spamming the same message everywhere too, it's unfortunate. I can't find his profile, is it banned or something?

", "rawMarkdown": "I saw him spamming the same message everywhere too, it's unfortunate. I can't find his profile, is it banned or something?", "votes": null }, { "id": "1458635", "postDate": "08/08/2021 01:51:09", "content": "

In my opinion, because his notebook is listed on the leaderboard. In the competition, his notebook is a public notebook with the second-highest ranking among all public notebooks. I gave you an upvote and thank you for your sharing. I am looking for a notebook using TensorFlow because my Macbook with m1 chip doesn't support PyTorch with GPU. I will cite your notebook in my forked notebook :)

", "rawMarkdown": "In my opinion, because his notebook is listed on the leaderboard. In the competition, his notebook is a public notebook with the second-highest ranking among all public notebooks. I gave you an upvote and thank you for your sharing. I am looking for a notebook using TensorFlow because my Macbook with m1 chip doesn't support PyTorch with GPU. I will cite your notebook in my forked notebook :)", "votes": null }, { "id": "1458690", "postDate": "08/08/2021 03:12:34", "content": "

Thank you for upvoting

", "rawMarkdown": "Thank you for upvoting", "votes": null }, { "id": "1459315", "postDate": "08/08/2021 10:14:36", "content": "

I tried to use AUC as a metric, but I failed. I saw you encoded the original dataset to tfdec format. Did you encode all images or just the images in the FLAIR folder? Have you labeled each image by the label of the patient? It seems like your dataset treats each image as an independent instance, so the CNN model just predicts a probability for each image, then averages the probability of images related to the patient. It doesn't make much sense to me, but I quite understand why your solution got a very well score.
\nI assume each patient with a tumor has a lot of images showing the tumor. On the other hand, some images don't show the tumor at all. The tumor should only show in part of the series in the images. If we labeled the tumor for each image, the images' labels should be like [0, 0, 0, 0… 1, 1, 1…, 1, 0] for a patient. If you just label images of ill patients with all 1s, and healthy patients with all 0s, it will give you a score lower than it expects.
\nTo improve the model, I think we can generate a table that labels each image. I want to manually separate images for positive patients. I noticed the positive patients have a white tumor-like mask at the upper right, we can find a rough beginning and end time that showing the tumor-like mask in the MRI videos and label those images as 1s, label the images before beginning, and after endpoints as 0s.

", "rawMarkdown": "I tried to use AUC as a metric, but I failed. I saw you encoded the original dataset to tfdec format. Did you encode all images or just the images in the FLAIR folder? Have you labeled each image by the label of the patient? It seems like your dataset treats each image as an independent instance, so the CNN model just predicts a probability for each image, then averages the probability of images related to the patient. It doesn't make much sense to me, but I quite understand why your solution got a very well score.\nI assume each patient with a tumor has a lot of images showing the tumor. On the other hand, some images don't show the tumor at all. The tumor should only show in part of the series in the images. If we labeled the tumor for each image, the images' labels should be like [0, 0, 0, 0... 1, 1, 1..., 1, 0] for a patient. If you just label images of ill patients with all 1s, and healthy patients with all 0s, it will give you a score lower than it expects. \nTo improve the model, I think we can generate a table that labels each image. I want to manually separate images for positive patients. I noticed the positive patients have a white tumor-like mask at the upper right, we can find a rough beginning and end time that showing the tumor-like mask in the MRI videos and label those images as 1s, label the images before beginning, and after endpoints as 0s.", "votes": null }, { "id": "1460570", "postDate": "08/09/2021 00:41:35", "content": "

Hi, I encoder al Images to tfrecords. Your approach Sounds good. Can you give an example for making the mask

", "rawMarkdown": "Hi, I encoder al Images to tfrecords. Your approach Sounds good. Can you give an example for making the mask", "votes": null }, { "id": "1461134", "postDate": "08/09/2021 08:23:35", "content": "

Thanks for upvoting

", "rawMarkdown": "Thanks for upvoting", "votes": null }, { "id": "1461210", "postDate": "08/09/2021 09:16:01", "content": "

For now, I just sorted filenames according to their number, then slide the images from the range 0.3 ~ 0.7 quartiles for each patient. After doing this, I noticed an obvious increasing trend of accuracy. The accuracy is improved from 0.5~0.6 to 0.8. But the AUC doesn't perform very well. I just have completed my modeling from scratch and finished the first test. I want to do a manual labeling work later this week. I will first label a few examples (~50) to see if the scores increase.

", "rawMarkdown": "For now, I just sorted filenames according to their number, then slide the images from the range 0.3 ~ 0.7 quartiles for each patient. After doing this, I noticed an obvious increasing trend of accuracy. The accuracy is improved from 0.5~0.6 to 0.8. But the AUC doesn't perform very well. I just have completed my modeling from scratch and finished the first test. I want to do a manual labeling work later this week. I will first label a few examples (~50) to see if the scores increase.", "votes": null }, { "id": "1461569", "postDate": "08/09/2021 13:13:42", "content": "

It's turn out that I introduced imbalanced data points into the dataset. Because my current datasets have 80% instances are nagtive, so it shows 80% accurary by random prediction😂.

", "rawMarkdown": "It's turn out that I introduced imbalanced data points into the dataset. Because my current datasets have 80% instances are nagtive, so it shows 80% accurary by random prediction😂.", "votes": null } ], "comments": [ { "id": 1402607, "author_name": "relurelu", "author_url": "", "post_date": "07/28/2021 10:55:47", "content": "

I think that posts were no respect for the original post. Thanks Luca !

", "votes": null, "replies": [] }, { "id": 1402636, "author_name": "carlmcbrideellis", "author_url": "", "post_date": "07/28/2021 11:39:46", "content": "

Dear @lucamtb

\n

On kaggle there is certainly a Matthew effect where higher ranked kagglers will procure, for better or for worse, more attention. However, in this case I would suggest the principal motive is the slight improvement in the LB score, which attracts some competition participants like moths to a flame. Indeed many such participants in competitions who do not undertake their own work are often caught out towards the end of many competitions by blindly submitting high scoring blended ensembles that turn out to be horrendously overfitted.
\nIt is certainly poor form to not have cited your original notebook when forking, more so given that the only difference was the slight change of one parameter.

\n

I hope you get the upvotes you deserve.

\n

All the best,
\ncarl

", "votes": null, "replies": [] }, { "id": 1403244, "author_name": "fazlencodes", "author_url": "", "post_date": "07/29/2021 00:34:46", "content": "

I have noticed that too.

", "votes": null, "replies": [] }, { "id": 1403467, "author_name": "ahmetcalis", "author_url": "", "post_date": "07/29/2021 06:34:59", "content": "

It happens most of time. At least you need to reference if you took some codes or inspired from another notebook. I hope people will learn it sooner or later.

", "votes": null, "replies": [] }, { "id": 1404555, "author_name": "shaz13", "author_url": "", "post_date": "07/30/2021 03:19:48", "content": "

I understand that the users filter notebooks on the score from the Code section and might have upvoted @blade001 's notebook from @lucamtb 's original work. The community would highly appreciate if the attribution of the original code author is added in the copied versions for better credibility.

\n

On the case of votes being cast - I think this is a noticeable issue with Kaggle as of now. Votes of notebooks don't pass to the original authors of the code work but that instance of the notebook.

", "votes": null, "replies": [] }, { "id": 1404585, "author_name": "prodzar", "author_url": "", "post_date": "07/30/2021 04:33:57", "content": "

That's sad and disappointing, but certainly happens, you should contact the person who fork your notebook and ask for the citation.

", "votes": null, "replies": [] }, { "id": 1404647, "author_name": "elenaeb", "author_url": "", "post_date": "07/30/2021 05:59:52", "content": "

I believe in fostering collaboration and creating a place where we (as fellow ML life-long learners) can be great support for one another. Sharing my own personal rule, in the hopes that it will inspire others to do so as well.

\n

a) if you find content that has captured your attention, or you've learned something new, or was worth your while in some form or another, or you became captured reading it -- upvote it. because ultimately, that content contributed something to you even if in the smallest minutiae

\n

b) if a notebook is useful in some manner or reason, not even to mention if it is one you're going to \"fork\", then definitely upvote it (not just take it for granted that it says \"copied from…\" but to actually give that person credit where credit is due; afterall you came across their notebook and used it).

\n

Perhaps I am verbose about this, more so at the moment, because I happen to have stumbled across many threads that seem to present information that could truly jeopardize what this platform's reputation is known to stand for (learning, collaboration, knowledge sharing), and the reason why I have joined.

\n

Look forward to learning from all of you!

", "votes": null, "replies": [] }, { "id": 1405030, "author_name": "dschettler8845", "author_url": "", "post_date": "07/30/2021 12:36:19", "content": "

If the new post doesn't add any value and doesn't give attribution to the original post you should report it.

\n

Simple as that.

\n

ps: I gave you an upvote so maybe that'll help.

", "votes": null, "replies": [ { "id": 1461134, "author_name": "lucamtb", "author_url": "", "post_date": "08/09/2021 08:23:35", "content": "

Thanks for upvoting

", "votes": null, "replies": [] } ] }, { "id": 1406310, "author_name": "boneacrabonjac", "author_url": "", "post_date": "07/31/2021 16:35:04", "content": "

Well out of his three notebooks he copied two and got gold medals for that…

", "votes": null, "replies": [] }, { "id": 1408336, "author_name": "furcifer", "author_url": "", "post_date": "08/02/2021 12:23:31", "content": "

I saw him spamming the same message everywhere too, it's unfortunate. I can't find his profile, is it banned or something?

", "votes": null, "replies": [] }, { "id": 1458635, "author_name": "bcghost", "author_url": "", "post_date": "08/08/2021 01:51:09", "content": "

In my opinion, because his notebook is listed on the leaderboard. In the competition, his notebook is a public notebook with the second-highest ranking among all public notebooks. I gave you an upvote and thank you for your sharing. I am looking for a notebook using TensorFlow because my Macbook with m1 chip doesn't support PyTorch with GPU. I will cite your notebook in my forked notebook :)

", "votes": null, "replies": [ { "id": 1458690, "author_name": "lucamtb", "author_url": "", "post_date": "08/08/2021 03:12:34", "content": "

Thank you for upvoting

", "votes": null, "replies": [] }, { "id": 1459315, "author_name": "bcghost", "author_url": "", "post_date": "08/08/2021 10:14:36", "content": "

I tried to use AUC as a metric, but I failed. I saw you encoded the original dataset to tfdec format. Did you encode all images or just the images in the FLAIR folder? Have you labeled each image by the label of the patient? It seems like your dataset treats each image as an independent instance, so the CNN model just predicts a probability for each image, then averages the probability of images related to the patient. It doesn't make much sense to me, but I quite understand why your solution got a very well score.
\nI assume each patient with a tumor has a lot of images showing the tumor. On the other hand, some images don't show the tumor at all. The tumor should only show in part of the series in the images. If we labeled the tumor for each image, the images' labels should be like [0, 0, 0, 0… 1, 1, 1…, 1, 0] for a patient. If you just label images of ill patients with all 1s, and healthy patients with all 0s, it will give you a score lower than it expects.
\nTo improve the model, I think we can generate a table that labels each image. I want to manually separate images for positive patients. I noticed the positive patients have a white tumor-like mask at the upper right, we can find a rough beginning and end time that showing the tumor-like mask in the MRI videos and label those images as 1s, label the images before beginning, and after endpoints as 0s.

", "votes": null, "replies": [] }, { "id": 1460570, "author_name": "lucamtb", "author_url": "", "post_date": "08/09/2021 00:41:35", "content": "

Hi, I encoder al Images to tfrecords. Your approach Sounds good. Can you give an example for making the mask

", "votes": null, "replies": [] }, { "id": 1461210, "author_name": "bcghost", "author_url": "", "post_date": "08/09/2021 09:16:01", "content": "

For now, I just sorted filenames according to their number, then slide the images from the range 0.3 ~ 0.7 quartiles for each patient. After doing this, I noticed an obvious increasing trend of accuracy. The accuracy is improved from 0.5~0.6 to 0.8. But the AUC doesn't perform very well. I just have completed my modeling from scratch and finished the first test. I want to do a manual labeling work later this week. I will first label a few examples (~50) to see if the scores increase.

", "votes": null, "replies": [] }, { "id": 1461569, "author_name": "bcghost", "author_url": "", "post_date": "08/09/2021 13:13:42", "content": "

It's turn out that I introduced imbalanced data points into the dataset. Because my current datasets have 80% instances are nagtive, so it shows 80% accurary by random prediction😂.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1402600": "Hi,\n\nI am asking myself why this notebook:\n\n[Brain Tumor](https://www.kaggle.com/blade001/brain-tumor-code)\n\ngot 26 upvote so far\n\nand this notebook:\n\n[Brain Tumor very basic inference](https://www.kaggle.com/lucamtb/brain-tumor-very-basice-inference)\n\ngot 10 upvote so far, \n\neven thought the first one was copied from the second one\n\nand the author of the first one, that copied not even gave a upvote", "1402607": "I think that posts were no respect for the original post. Thanks Luca !", "1402636": "Dear @lucamtb \n\nOn kaggle there is certainly a [Matthew effect](https://en.wikipedia.org/wiki/Matthew_effect) where higher ranked kagglers will procure, for better or for worse, more attention. However, in this case I would suggest the principal motive is the slight improvement in the LB score, which attracts some competition participants like moths to a flame. Indeed many such participants in competitions who do not undertake their own work are often caught out towards the end of many competitions by blindly submitting high scoring blended ensembles that turn out to be horrendously overfitted.\nIt is certainly poor form to not have cited your original notebook when forking, more so given that the only difference was the slight change of one parameter. \n\nI hope you get the upvotes you deserve.\n\nAll the best,\ncarl", "1403244": "I have noticed that too.", "1403467": "It happens most of time. At least you need to reference if you took some codes or inspired from another notebook. I hope people will learn it sooner or later.", "1404555": "I understand that the users filter notebooks on the score from the Code section and might have upvoted @blade001 's notebook from @lucamtb 's original work. The community would highly appreciate if the attribution of the original code author is added in the copied versions for better credibility. \n\nOn the case of votes being cast - I think this is a noticeable issue with Kaggle as of now. Votes of notebooks don't pass to the original authors of the code work but that instance of the notebook.", "1404585": "That's sad and disappointing, but certainly happens, you should contact the person who fork your notebook and ask for the citation.", "1404647": "I believe in fostering collaboration and creating a place where we (as fellow ML life-long learners) can be great support for one another. Sharing my own personal rule, in the hopes that it will inspire others to do so as well.\n\na) if you find content that has captured your attention, or you've learned something new, or was worth your while in some form or another, or you became captured reading it -- upvote it. because ultimately, that content contributed something to you even if in the smallest minutiae\n\nb) if a notebook is useful in some manner or reason, not even to mention if it is one you're going to \"fork\", then definitely upvote it (not just take it for granted that it says \"copied from...\" but to actually give that person credit where credit is due; afterall you came across their notebook and used it).\n\nPerhaps I am verbose about this, more so at the moment, because I happen to have stumbled across many threads that seem to present information that could truly jeopardize what this platform's reputation is known to stand for (learning, collaboration, knowledge sharing), and the reason why I have joined. \n\nLook forward to learning from all of you!", "1405030": "If the new post doesn't add any value and doesn't give attribution to the original post you should report it. \n\nSimple as that.\n\nps: I gave you an upvote so maybe that'll help.", "1406310": "Well out of his three notebooks he copied two and got gold medals for that...", "1408336": "I saw him spamming the same message everywhere too, it's unfortunate. I can't find his profile, is it banned or something?", "1458635": "In my opinion, because his notebook is listed on the leaderboard. In the competition, his notebook is a public notebook with the second-highest ranking among all public notebooks. I gave you an upvote and thank you for your sharing. I am looking for a notebook using TensorFlow because my Macbook with m1 chip doesn't support PyTorch with GPU. I will cite your notebook in my forked notebook :)", "1458690": "Thank you for upvoting", "1459315": "I tried to use AUC as a metric, but I failed. I saw you encoded the original dataset to tfdec format. Did you encode all images or just the images in the FLAIR folder? Have you labeled each image by the label of the patient? It seems like your dataset treats each image as an independent instance, so the CNN model just predicts a probability for each image, then averages the probability of images related to the patient. It doesn't make much sense to me, but I quite understand why your solution got a very well score.\nI assume each patient with a tumor has a lot of images showing the tumor. On the other hand, some images don't show the tumor at all. The tumor should only show in part of the series in the images. If we labeled the tumor for each image, the images' labels should be like [0, 0, 0, 0... 1, 1, 1..., 1, 0] for a patient. If you just label images of ill patients with all 1s, and healthy patients with all 0s, it will give you a score lower than it expects. \nTo improve the model, I think we can generate a table that labels each image. I want to manually separate images for positive patients. I noticed the positive patients have a white tumor-like mask at the upper right, we can find a rough beginning and end time that showing the tumor-like mask in the MRI videos and label those images as 1s, label the images before beginning, and after endpoints as 0s.", "1460570": "Hi, I encoder al Images to tfrecords. Your approach Sounds good. Can you give an example for making the mask", "1461134": "Thanks for upvoting", "1461210": "For now, I just sorted filenames according to their number, then slide the images from the range 0.3 ~ 0.7 quartiles for each patient. After doing this, I noticed an obvious increasing trend of accuracy. The accuracy is improved from 0.5~0.6 to 0.8. But the AUC doesn't perform very well. I just have completed my modeling from scratch and finished the first test. I want to do a manual labeling work later this week. I will first label a few examples (~50) to see if the scores increase.", "1461569": "It's turn out that I introduced imbalanced data points into the dataset. Because my current datasets have 80% instances are nagtive, so it shows 80% accurary by random prediction😂." }, "source": "meta" }